Reproduce “VeloBins: Learning Velocity and Its Uncertainty via Bins and Error-Conditioned Gaussian Labels for Aerial Inertial Odometry” on OpenResearch

About this paper

Maulana Bisyir Azhari, Seungwook Lee, Donghun Han, Sung Jun Park, David Hyunchul Shim

The VeloBins method reformulates velocity estimation for aerial inertial odometry as a classification task over discretized bins, using error-conditioned Gaussian labels to supervise uncertainty. This framework improved trajectory estimation, reducing Absolute Trajectory Error by up to 53% compared to the strongest baseline, and delivered better-calibrated uncertainty for more consistent Extended Kalman Filter operation in dynamic aerial robot maneuvers.

Key insights

• Recasting continuous velocity regression as classification over bins inherently improves accuracy and trainability for highly dynamic systems like aerial robots, even with basic supervision.
• Explicitly supervising uncertainty with error-conditioned Gaussian labels, whose variance reflects the true error, leads to better-calibrated uncertainty estimates and enhanced filter consistency without directly optimizing an NLL loss.
• Deriving both the velocity mean and variance from a single bin distribution provides a unified, robust, and interpretable mechanism for predicting both the state and its confidence.

Summary by alphaXiv.